![Application of Haralick Texture Features in Brain [18F]-Florbetapir Positron Emission Tomography Without Reference Region Normalization](http://data.docslib.org/img/90ce2a396afa133c837bde031b8d8157-1.webp)
Journal name: Clinical Interventions in Aging Article Designation: Original Research Year: 2017 Volume: 12 Clinical Interventions in Aging Dovepress Running head verso: Campbell et al Running head recto: Application of HFs in brain [18F]-florbetapir PET open access to scientific and medical research DOI: http://dx.doi.org/10.2147/CIA.S143307
Open Access Full Text Article Original Research Application of Haralick texture features in brain [18F]-florbetapir positron emission tomography without reference region normalization
Desmond L Campbell1 Objectives: Semi-quantitative image analysis methods in Alzheimer’s Disease (AD) require Hakmook Kang2 normalization of positron emission tomography (PET) images. However, recent studies have Sepideh Shokouhi1 found variabilities associated with reference region selection of amyloid PET images. Haralick features (HFs) generated from the Gray Level Co-occurrence Matrix (GLCM) quantify spatial On behalf of The characteristics of amyloid PET radiotracer uptake without the need for intensity normalization. Alzheimer’s Disease The objective of this study is to calculate several HFs in different diagnostic groups and deter- Neuroimaging Initiative mine the group differences. 1Department of Radiology and Methods: All image and metadata were acquired through the Alzheimer’s Disease Neuroimaging Radiological Sciences, 2Department Initiative database. Subjects were grouped in three ways: by clinical diagnosis, by APOE e4 of Biostatistics, Vanderbilt University allele, and by Alzheimer’s Disease Assessment Scale-cognitive subscale (ADAS-Cog) score. Medical Center, Vanderbilt University Institute of Imaging Science, Nashville, Several GLCM matrices were calculated for different direction and distances (1–4 mm) from TN, USA multiple regions on PET images. The HFs, contrast, correlation, dissimilarity, energy, entropy, and homogeneity, were calculated from these GLCMs. Wilcoxon tests and Student t-tests were performed on Haralick features and standardized uptake value ratio (SUVR) values, respectively, to determine group differences. In addition to statistical testing, receiver operat- ing characteristic (ROC) curves were generated to determine the discrimination performance of the selected regional HFs and the SUVR values. Results: Preliminary results from statistical testing indicate that HFs were capable of distin- guishing groups at baseline and follow-up (false discovery rate corrected p,0.05) in particular regions at much higher occurrences than SUVR (81 of 252). Conversely, we observed nearly no significant differences between all groups within ROIs at baseline or follow-up utilizing SUVR. From the ROC analysis, we found that the Energy and Entropy offered the best per- formance to distinguish Normal versus mild cognitive impairment and ADAS-Cog negative versus ADAS-Cog positive groups. Conclusion: These results suggest that this technique could improve subject stratification in AD drug trials and help to evaluate the disease progression and treatment effects longitudinally without the disadvantages associated with intensity normalization. Keywords: Haralick features, florbetapir, gray level co-occurrence matrix, energy, entropy
Correspondence: Sepideh Shokouhi Department of Radiology and Radiological Introduction Sciences, Vanderbilt University Medical Regional and voxel-based semi-quantitative image analysis methods in Alzheimer’s Center, Vanderbilt University Institute Disease (AD) research require normalization of positron emission tomography of Imaging Science, 1161 21st Avenue South, Medical Center North, AA-1105, (PET) images before data analysis. Most amyloid-PET (Aβ-PET) studies in AD Nashville, TN 37232-2310, USA use cerebellum as the reference region,1–4 however, recent research has found vari- Tel +1 631 513 0547 5–8 Email [email protected] abilities associated with the cerebellar normalization of amyloid PET images. submit your manuscript | www.dovepress.com Clinical Interventions in Aging 2017:12 2077–2086 2077 Dovepress © 2017 Campbell et al. This work is published and licensed by Dove Medical Press Limited. The full terms of this license are available at https://www.dovepress.com/terms.php and incorporate the Creative Commons Attribution – Non Commercial (unported, v3.0) License (http://creativecommons.org/licenses/by-nc/3.0/). By accessing the work you http://dx.doi.org/10.2147/CIA.S143307 hereby accept the Terms. Non-commercial uses of the work are permitted without any further permission from Dove Medical Press Limited, provided the work is properly attributed. For permission for commercial use of this work, please see paragraphs 4.2 and 5 of our Terms (https://www.dovepress.com/terms.php). Campbell et al Dovepress
White matter (WM) can be used for Aβ-PET normalization Subjects to reduce longitudinal variability,5 increase association to The key eligibility criteria of ADNI subject recruitment are 7 8 clinical decline and cerebrospinal fluid levels of Aβ1–42 as available on the ADNI website (http://www.adni-info.org). well enhance discrimination power between subject groups.6 Briefly, enrolled ADNI subjects were between the ages of 55 However, other recent studies have found impaired amyloid and 90, had a study partner able to provide an independent PET radiotracer uptake in damaged areas of WM.9 WM evaluation of functioning, and spoke either English or injury is common in aging and dementia.10–13 Therefore, Spanish. All subjects gave written, informed consent before the utility of WM as a reference region may depend on its participation through their local Institutional Review Board. structural and functional integrity, which can vary among All images and metadata for 30 subjects (17 male, 13 female) individual subjects. were acquired from the ADNI database. Selected subjects for Based on these findings, we have identified that the this study had undergone longitudinal [18F]-florbetapir PET amyloid PET normalization process poses a critical challenge scans with concurrent T1-weighted MRIs at baseline and for semi-quantitative PET studies in AD. In this work, we 24-month follow-up. In addition, scores from the neurop- evaluate a different semi-quantitative PET approach that does sychological assessment Alzheimer’s Disease Assessment not require the intensity normalization of Aβ-PET activity to Scale-cognitive subscale (ADAS-Cog)15 were collected and a reference region activity. This method is based on Haralick time-matched to the nearest Aβ-PET acquisition. Table 1 features (HFs) that can be calculated from the Gray Level illustrates the demographic and clinical data for selected Co-occurrence Matrices (GLCMs) of Aβ-PET images to subjects. The 30 subjects were stratified three ways: by provide a statistical description of the spatial characteristics clinical diagnosis (15 normal versus 15 MCI), by APOE e4 of amyloid PET.14 The objective of this study is to determine allele carrier status (14 non-carriers versus 16 carriers), and whether higher-order texture-based features derived from by ADAS-Cog score defined as positive by a threshold$ 9 at the GLCMs of [18F]-Florbetapir-PET images can identify baseline or follow-up (16 negative versus 14 positive). statistically and clinically significant differences between subject groups without using a reference region for PET [18F]-florbetapir PET data acquisition intensity normalization. All participating sites acquired the [18F]-florbetapir16 PET imaging according to standardized ADNI protocols (http:// Methods adni.loni.usc.edu/methods/pet-analysis/pre-processing/). Alzheimer’s disease neuroimaging Serial PET images for our study were acquired at 13 different initiative centers, and the data were obtained following the dynamic The data presented in this study were acquired from the protocol of and preprocessed by ADNI. Dynamic scans were Alzheimer’s Disease Neuroimaging Initiative (ADNI) data- 30-min 6-frame scans acquired 30–60 min after injection. base (adni.loni.usc.edu). The ADNI was launched in 2003 All frames were coregistered to the first frame, summed, and as a public-private partnership, led by Principal Investigator averaged to create a single image volume. Michael W Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic resonance imaging [18F]-florbetapir PET pre-processing (MRI), PET, other biological markers, and clinical and We employed SPM12 (Wellcome Department of Cognitive neuropsychological assessment can be combined to measure Neurology) for image pre-processing. The gray matter frac- the progression of mild cognitive impairment (MCI) and tions of 7 anatomic regions were extracted and utilized for early AD. ROI analysis: Combined precuneus/posterior cingulate cortex
Table 1 Demographic Alzheimer’s disease neuroimaging initiative subject data for this study Stratification Normal MCI APOE non APOE ADAS-Cog ADAS-Cog carrier carrier negative positive Number of subjects 15 15 14 16 16 14 Female (male) 8 (7) 5 (10) 8 (6) 5 (11) 6 (10) 7 (7) Age (mean ± SD) 74.7±4.3 73.6±8.7 77.4±7.4 74.6±6.7 74.7±4.9 77.2±9.0 ADAS-Cog (mean ± SD) 6.7±4.2 12.7±7.7 9.9±8.1 9.6±5.7 5.4±2.3 14.6±7.0 Abbreviations: ADAS-Cog, Alzheimer’s Disease Assessment Scale-cognitive subscale; MCI, mild cognitive impairment.
2078 submit your manuscript | www.dovepress.com Clinical Interventions in Aging 2017:12 Dovepress Dovepress Application of HFs in brain [18F]-florbetapir PET
(PCC), left & right frontal lobes (LFL & RFL), left & right pixel values between a reference and neighboring pixels for parietal lobes (LPL & RPL), left & right temporal lobes a particular distance and direction. (LTL & RTL), and cerebellum for reference region normal- We used the graycomatrix() function in MATLAB (The ization. For each subject, baseline T1-weighted MR image MathWorks, Inc., Natick, Massachusetts, USA) to generate volumes defined the anatomical ROIs in native space using a all GLCMs. The GLCM was calculated by specifying an maximum probability tissue labels derived from the Medical offset or displacement vector and counting all pairs of Image Computing and Computer Assisted Interventions Con- voxels separated by this offset having intensity levels i and j. ference 2012 Grand Challenge and Workshop on Multi-Atlas By default, the graycomatrix function in MATLAB calcu- Labeling and provided by Neuromorphometrics, Inc.17–19 lates the GLCM based on the horizontal proximity of the pixels: (0 1). That is the pixel next to the pixel of interest Gray level co-occurrence matrix on the same row. However, other voxel spatial relationships The GLCM texture analysis is a method for evaluating can be specified by using the “Offsets” parameter as input. higher-order statistical methods in two-dimensional (2D) This approach allows for multiple GLCMs with different images. First-order statistics metrics from histograms evalu- directions and distances to be generated from a single image. ate only pixel intensities, while GLCM assesses their spatial Figure 1 highlights this process for creating representa- associations. To generate the GLCM, pixel intensities within tive GLCMs from a sample image. For each axial slice, 2D images were first discretized. The GLCM was then cal- 34 GLCMs were calculated for different offsets (Figure 1A). culated by tabulating the occurrences of a combination of Then, an average GLCM (GLCM_avg) was calculated from
$ %
± ± ± ± ± ± ,PDJHYHUWLFDOD[LV ±
± HUWLFDOGLVWDQFH YR[HOV
9 ,PDJHKRUL]RQWDOD[LV ±
± +RUL]RQWDOGLVWDQFH YR[HOV & 9R[HOYDOXH ' 9R[HOYDOXH
R[HOYDOXH R[HOYDOXH 9 9
Figure 1 (A) The coordinates of 34 offsets with respect to (0, 0). These 34 different offsets were used to calculate 33 gray level co-occurrence matrices (GLCMs) for each axial slice of a regional positron emission tomography image. (B) A representative image where two sample voxel pairs for two different offsets, (0°, distance =1) and (135°, distance =1) are highlighted. The discrete image voxel values are 1, 2, 3, and 4. (C) GLCM of image (B) for (0°, distance =1). The x and y axis represent the ordered (from low to high) voxel values of image (B). The z values represent the number of occurrences of a specific pair of voxel values (eg, 3 and 2) for a given offset (0°, distance =1). (D) GLCM of image (B) for (135°, distance =1). The x and y axis represent the ordered (from low to high) voxel values of image (B). The z values represent the number of occurrences of a specific pair of voxel values (eg, 3 and 2) for a given offset (135°, distance =1).
Clinical Interventions in Aging 2017:12 submit your manuscript | www.dovepress.com 2079 Dovepress Campbell et al Dovepress these individual GLCMs. The HFs were calculated from of these GLCMs was used to calculate HFs. As a standard each GLCM_avg, and then the features were averaged over method for comparison, the semi-quantitative regional SUVR multiple axial slices. The calculation of the GLCM was values were calculated by normalizing the images to the performed in 2D mode because the graycomatrix() function cerebellar activity. requires 2D images. Statistical analysis Haralick features (HFs) Statistical testing was performed to determine if the SUVR Following six HFs20,21 were calculated from the GLCM_avg measurements and the selected HFs provide significant matrices: differences between stratified groups (normal versus MCI; APOE-ε4 non-carriers versus carriers; ADAS-Cog negative 2 Contrast =−ij× pi(,j) versus positive) at baseline and 24-month follow-up for each ∑ ij, ROI structure. Unpaired, nonparametric Wilcoxon statistical ()ij−−µµ()pi(,j) tests were employed to compare HFs, while unpaired, Correlation = ij; ∑ ij, Student’s t-tests were utilized for comparing SUVR mea- σσij surements. A 95% CI was set, correlating to p-values ,0.05 µ =×ip(,ij); i ∑ i required for significant differences between stratified groups. σ = (i −−×µ )(2 pi,)j Between 3 stratifications of groups, 7 ROIs, and two time i ∑ i i points, 42 Student’s t-tests for SUVR were conducted, and Dissimilarity =−ijjp× (,i ) using six different HFs, 252 total Wilcoxon tests were per- ∑ ij, formed. For a group of ROIs and HFs, we controlled False Energy = p()ij, 2 Discovery Rate (FDR) at 0.05 level to deal with multiple ∑ ij, comparisons. Entropy =− pp()ij,I×+n ()ij, ζζ; = 10−7 In addition to statistical testing, receiver operating char- ∑ ij, acteristic (ROC) curves were generated to determine the 1 discrimination performance of the selected regional HFs and Homogeneity = × p()ij, ∑ ij, 1 ij +− the SUVR values. Groups were stratified as described previ- ously except that data from HFs and SUVR from baseline where i and j are the pixel indices and p(i, j) represents the and 24-month follow-up were combined. The Area under the GLCM pixel intensities. Briefly describing the HFs, Contrast Curve (AUC) was used to quantitatively characterize each measures the local variations in the GLCM. Correlation generated ROC curve. measures the linear dependence of the pixel intensity related to its position in the image. Dissimilarity is akin to Contrast Results with measuring local variations, however, it has a linear Gray Level Co-occurrence matrices were generated for slices dependent off-diagonal of the GLCM. Energy is associated through predefined brain structures. Shown in Figure 2 are with local homogeneity in the image. Entropy, similar to a representative slice of the discretized PCC from a normal how it is defined in physics, measures the spatial disorder and MCI subject used in the GLCM creation. Qualitatively, in the GLCM. Finally, Homogeneity measures how similar the PCC of the normal subject has a lower intensity than the the pixel intensities of the GLCM are, and thus, is akin to PCC of the MCI. Accompanying each PCC slice are their a measure of uniformity. It is our hypothesis that HFs may GLCMs, averaged over our distance and direction range. be sensitive to changes in spatial distribution and amyloid In the representative case shown, we observe a greater degree activity among different subject groups. of clustering in the normal GLCM versus the MCI GLCM, For this study, GLCMs were generated from the largest which exhibits a larger spread along its diagonal. Comparing five axial slices through the PCC, LFL, RFL, LPL, RPL, the GLCM of normal and MCI subjects, we observed GLCMs LTL, and RTL. Before generating the GLCM, each axial with more occurrences at greater discretized intensities for slice was discretized into 128 intensity bins. For each slice, MCI subjects. GLCMs were generated for each distance between 1 and Results from the Student’s t- and Wilcoxon tests are 4 pixels at angles between 0 and 165 degrees. The average displayed in Tables 2–4 with boxplots comparing the HFs
2080 submit your manuscript | www.dovepress.com Clinical Interventions in Aging 2017:12 Dovepress Dovepress Application of HFs in brain [18F]-florbetapir PET
$ % D[LV D[LV 3(7LPDJHYHUWLFDO 3(7LPDJHYHUWLFDO 3(7LPDJHKRUL]RQWDOD[LV 3(7LPDJHKRUL]RQWDOD[LV & ' R[HOYDOXHRI R[HOYDOXHRI
9
9
GLVFUHWL]HG3(7LPDJH GLVFUHWL]HG3(7LPDJH 9R[HOYDOXHRI 9R[HOYDOXHRI GLVFUHWL]HG3(7LPDJH GLVFUHWL]HG3(7LPDJH
Figure 2 Representative axial slices of Aβ-PET images in PCC from (A) single normal subject and (B) single MCI subject. Accompanying each PCC slice are their gray level co-occurrence matrices, normal (C) and MCI (D), averaged over a range of distances and directions. Abbreviations: MCI, mild cognitive impairment; PCC, posterior cingulate cortex; PET, positron emission tomography. and SUVRs for the precuneus and PCC shown in Figure 3. LFL). For the APOE and ADAS-Cog stratified groups, fewer Overall, the HFs were capable of distinguishing groups at HFs were capable of distinguishing groups in specific brain baseline and follow-up (at FDR =0.05 level) in particular regions. However, we found significant differences in these regions. In total, 124 out of 252 Wilcoxon tests showed sig- groups using Energy and Entropy in the PCC at baseline for nificant differences between stratified groups. The majority ADAS-Cog in Table 4 and follow-up for APOE in Table 3. of these differences were exhibited in the subjects grouped Conversely, we observed no significant differences between by diagnosis at both baseline and 24-month follow-up. Spe- all groups and ROIs at baseline or follow-up utilizing SUVR cifically, Energy and Entropy were capable of distinguishing (0 of 42 performed tests). normal and MCI subjects in all brain regions at baseline and ROC curves for the PCC and all AUC values are dis- nearly all structures at follow-up (except for Energy in the played in Figure 4. The ROC AUCs are presented in Table 5,
Table 2 FDR-adjusted p-values of HFs and SUVR group difference (N, MCI) at baseline and follow-up Feature Normal vs MCI at baseline Normal vs MCI at follow-up PCC LFL RFL LPL RPL LTL RTL PCC LFL RFL LPL RPL LTL RTL Contrast 0.0506 0.0163 0.0118 0.0245 0.0049 0.1576 0.4649 0.3589 0.0391 0.0871 0.0259 0.0044 0.8188 0.981 Correlation 0.0353 0.0506 0.0317 0.0689 0.0868 0.0049 0.0118 0.0044 0.0608 0.1917 0.8983 0.1141 0.174 0.1118 Dissimilarity 0.0749 0.0184 0.0245 0.0298 0.0049 0.2251 0.3836 0.3907 0.0397 0.0572 0.0229 0.0044 0.6403 0.7915 Energy 0.0049 0.0406 0.0082 0.0049 0.0049 0.0381 0.0381 0.0044 0.0259 0.0082 0.0044 0.0044 0.3589 0.0381 Entropy 0.0049 0.0698 0.014 0.0049 0.0049 0.014 0.0082 0.0044 0.0423 0.0151 0.0044 0.0044 0.0622 0.0082 Homogeneity 0.3099 0.0245 0.0385 0.0245 0.0049 0.1987 0.1566 0.5098 0.0533 0.0392 0.0175 0.0044 0.168 0.315 SUVR 0.385 0.6054 0.6338 0.74 0.74 0.7319 0.5245 0.2688 0.4134 0.4163 0.6292 0.5994 0.4709 0.3055 Note: Controlling for FDR at 0.05, the adjusted p-values smaller than 0.05 (blue color) indicate significant group differences. Abbreviations: FDR, false discovery rate; HFs, Haralick features; LFL, left frontal lobe; LPL, left parietal lobe; LTL, left temporal lobe; MCI, mild cognitive impairment; PCC, posterior cingulate cortex; RFL, right frontal lobe; RPL, right parietal lobe; RTL, right temporal lobe; SUVR, standardized uptake value ratio.
Clinical Interventions in Aging 2017:12 submit your manuscript | www.dovepress.com 2081 Dovepress Campbell et al Dovepress
Table 3 FDR-adjusted p-values of HFs and SUVR group difference (APOE-negative, APOE-positive) at baseline and follow-up Feature APOE-E4 carriers vs non-carriers at baseline APOE-E4 carriers vs non-carriers at follow-up PCC LFL RFL LPL RPL LTL RTL PCC LFL RFL LPL RPL LTL RTL Contrast 0.3185 0.5641 0.2646 0.5488 0.5641 0.5042 0.0926 0.28 0.385 0.0716 0.5278 0.2096 0.4883 0.968 Correlation 0.0858 0.2321 0.1895 0.273 0.147 0.2321 0.3308 0.6995 0.5278 0.0098 0.2838 0.0098 0.0882 0.0612 Dissimilarity 0.3185 0.5488 0.1895 0.4807 0.5673 0.4237 0.1029 0.1729 0.3524 0.0381 0.5278 0.1729 0.3739 0.9606 Energy 0.0858 0.5641 0.3383 0.3185 0.2321 0.0327 0.022 0.0392 0.8501 0.5278 0.8929 0.2838 0.0098 0.0098 Entropy 0.0858 0.5673 0.3383 0.3163 0.2321 0.049 0.022 0.049 0.8542 0.5278 0.8501 0.3018 0.0098 0.0122 Homogeneity 0.4083 0.4805 0.049 0.3185 0.8014 0.3308 0.1069 0.0122 0.2254 0.0122 0.5814 0.0882 0.2254 0.6995 SUVR 0.5837 0.4805 0.4083 0.918 0.7079 0.3379 0.1895 0.8501 0.6486 0.573 0.9606 0.8969 0.5278 0.2838 Note: Controlling for FDR at 0.05, the adjusted p-values smaller than 0.05 (blue color) indicate significant group differences. Abbreviations: FDR, false discovery rate; HFs, Haralick features; LFL, left frontal lobe; LPL, left parietal lobe; LTL, left temporal lobe; PCC, posterior cingulate cortex; RFL, right frontal lobe; RPL, right parietal lobe; RTL, right temporal lobe; SUVR, standardized uptake value ratio. respectively. From Figure 4A and C, we observed that the were classified with respect to their clinical diagnosis, Energy and Entropy offered the best performance to distin- APOE-ε4 status, and ADAS-Cog score to make a comparison guish Normal versus MCI and ADAS-Cog negative versus between Haralick texture features and the conventional ADAS-Cog positive groups. Conversely, SUVR and Entropy SUVR approach. provided the best methods APOE-ε4 carriers versus non- The results of the statistical and diagnostic testing revealed carriers, as viewed in Figure 4B, while Energy exhibited that the HFs Energy and Entropy provide the best descriptor lower performance. From Table 5, we observed that several for distinguishing normal and MCI subjects across all brain metrics were best for differentiating stratified groups. For the structures of interest. There was statistically significant sepa- classifications based on clinical diagnosis and ADAS-Cog, ration in Energy and Entropy between the normal and MCI Energy gave the best performance across all ROIs. For subject populations in the majority of ROIs at baseline and the APOE-ε4 classification, Entropy performed the best, 24-month follow-up. Energy and Entropy also offered the followed by SUVR and Dissimilarity as secondary metrics. largest AUC values for accurately classifying subjects with Homogeneity ranked as the second best metric for the ADAS- respect to their clinical diagnosis (Normal versus MCI) or Cog classification. Several other HFs exhibited AUC values ADAS-Cog status (negative versus positive). close to 0.5, considerably not better than random guessing. An interesting consequence observed from the results is the apparent disconnect between the statistical testing and Discussion ROC diagnostic performance. Results from the Student’s The objective of this study was to determine whether higher- t-tests determined that no significant differences were exhib- order texture-based features derived from the GLCMs of ited in SUVR for each stratified group in all ROIs. However, [18F]-Florbetapir-PET images can identify statistically sig- ROC curve analysis verified SUVR’s capability to distin- nificant differences between subject groups without using a guish populations, specifically in the RTL. This diagnostic reference region for PET intensity normalization. Subjects capability also extended to all brain regions for the APOE
Table 4 FDR-adjusted p-values of HFs and SUVR group difference (ADAS-Cog-negative, ADAS-Cog-positive) at baseline and follow-up Feature ADAS-Cog negative vs positive at baseline ADAS-Cog negative vs positive at follow-up PCC LFL RFL LPL RPL LTL RTL PCC LFL RFL LPL RPL LTL RTL Contrast 0.1994 0.1994 0.0754 0.1315 0.0686 0.1315 0.1994 0.2554 0.3369 0.294 0.1102 0.091 0.091 0.127 Correlation 0.6136 0.7809 0.4923 0.3127 0.3416 0.686 0.3127 0.091 0.4477 0.5484 0.2292 0.1102 0.1903 0.1102 Dissimilarity 0.2282 0.1994 0.0754 0.1315 0.07 0.1446 0.1994 0.343 0.3237 0.2554 0.1102 0.091 0.091 0.1823 Energy 0.049 0.446 0.1994 0.1315 0.0754 0.049 0.2282 0.1102 0.2292 0.1449 0.1102 0.1102 0.1102 0.2292 Entropy 0.049 0.4139 0.1994 0.1315 0.0754 0.049 0.1994 0.1102 0.2292 0.1213 0.1102 0.1102 0.091 0.1674 Homogeneity 0.3416 0.1994 0.0754 0.0754 0.07 0.1315 0.2282 0.6064 0.3205 0.2292 0.1102 0.091 0.1102 0.3205 SUVR 0.3668 0.4957 0.4923 0.4923 0.6032 0.818 0.5891 0.3205 0.357 0.3292 0.3829 0.4676 0.666 0.453 Note: Controlling for FDR at 0.05, the adjusted p-values smaller than 0.05 (blue color) indicate significant group differences. Abbreviations: ADAS-Cog, Alzheimer’s Disease Assessment Scale-cognitive subscale; FDR, false discovery rate; HFs, Haralick features; LFL, left frontal lobe; LPL, left parietal lobe; LTL, left temporal lobe; PCC, posterior cingulate cortex; RFL, right frontal lobe; RPL, right parietal lobe; RTL, right temporal lobe; SUVR, standardized uptake value ratio.
2082 submit your manuscript | www.dovepress.com Clinical Interventions in Aging 2017:12 Dovepress Dovepress Application of HFs in brain [18F]-florbetapir PET
$ % S S S S H
&RQWUDVW &RUUHODWLRQ 1RUPDO 0&, $32( $32(H $'$6&RJ $'$6&RJ 1RUPDO 0&, $32( $32(H $'$6&RJ $'$6&RJ HH QHJDWLYH SRVLWLYH HH QHJDWLYH SRVLWLYH
& ' S H± S (QHUJ\
'LVVLPLODULW\
1RUPDO 0&, $32( $32(H $'$6&RJ $'$6&RJ 1RUPDO 0&, $32( $32(H $'$6&RJ $'$6&RJ HH QHJDWLYH SRVLWLYH HH QHJDWLYH SRVLWLYH
( S H± S )
(QWURS\
+RPRJHQHLW\
1RUPDO 0&, $32( $32(H $'$6&RJ $'$6&RJ 1RUPDO 0&, $32( $32(H $'$6&RJ $'$6&RJ HH QHJDWLYH SRVLWLYH HH QHJDWLYH SRVLWLYH *
6895
1RUPDO 0&, 1RQFDUULHU &DUULHU $'$6&RJ $'$6&RJ QHJDWLYH SRVLWLYH
Figure 3 Boxplots of measured metrics for the combined precuneus and posterior cingulate cortex. Notes: (A) Contrast, (B) Correlation, (C) Dissimilarity, (D) Energy, (E) Entropy, (F) Homogeneity, (G) SUVR. Metrics with statistically significant differences are shown with corresponding p-values ,0.05. Abbreviations: MCI, mild cognitive impairment; SUVR, standardized uptake value ratio. stratified group, with AUC values. 0.5, including the PCC, is observed in Aβ-PET images.23 These spatial changes can as seen in Figure 4B. Recall that AUC can be interpreted as affect the local disorder, homogeneity, variation, and similarity diagnosis accuracy as a percentage, where 0.5 represents of pixel intensities, which can be captured by HFs. 50% accuracy. Even though SUVR appears to be a sensitive One of the limitations of this study is the small sample metric for differentiating APOE populations, other texture size, which limits the interpretation regarding diagnostic features perform better with respect to the two other clas- capabilities. However, the overall objective was to compare sification schemes. a new image analysis method with a standard technique. The HFs are abstract mathematical models that are used in outcomes of our study indicate that, for the same sample size, many different applications. The innovative aspect of this study our new method shows more significant group differences was to utilize these features for characterizing abnormal Aβ than the SUVR values. pathology from non-normalized florbetapir-PET images. From early post-mortem studies,22 we know that over time the Conclusion progressive Aβ accumulation encompasses a greater extent of In this pilot study, a novel texture-based approach to assess- cerebral cortical laminae. The same pattern of spatial spread ing diagnostic statuses of amyloid burden was evaluated.
Clinical Interventions in Aging 2017:12 submit your manuscript | www.dovepress.com 2083 Dovepress Campbell et al Dovepress
$%
6HQVLWLYLW\ 6HQVLWLYLW\
±VSHFLILFLW\ ±VSHFLILFLW\
&
6HQVLWLYLW\
±VSHFLILFLW\ &RQWUDVW &RUUHODWLRQ 'LVVLPLODULW\ (QHUJ\ (QWURS\ +RPRJHQHLW\ 6895
Figure 4 Representative ROC curves characterizing the diagnostic capability of HFs and SUVR measurements within the combined precuneus and posterior cingulate cortex. Notes: ROC curves for the groups stratified by (A) Alzheimer’s disease neuroimaging initiative clinical diagnosis (N versus mild cognitive impairment), (B) APOE status, and (C) ADAS-Cog status. The HF Energy exhibits the greatest capability to distinguish subjects in the Alzheimer’s disease diagnosis and ADAS-Cog stratified groups. Abbreviations: ADAS-Cog, Alzheimer’s Disease Assessment Scale-cognitive subscale; HFs, Haralick features; ROC, receiver operating characteristic; SUVR, standardized uptake value ratio.
We demonstrated that HFs, specifically Energy, could be Noor Tantawy at Vanderbilt University Institute of Imaging used to effectively classify patients with respect to their Science for supportive discussions. Data collection and shar- clinical diagnosis, APOE-ε4 status, and ADAS-Cog score, ing for this project was funded by the Alzheimer’s Disease compared to SUVR. These results suggest that this technique Neuroimaging Initiative (ADNI) (National Institutes of has potentials to improve subject stratification in AD drug Health Grant U01 AG024904) and DOD ADNI (Department trials and help evaluate the disease progression and treatment of Defense award number W81XWH-12-2-0012). ADNI effects longitudinally without the potential biases associated is funded by the National Institute on Aging, the National with the reference region normalization. Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: Acknowledgments AbbVie, Alzheimer’s Association; Alzheimer’s Drug This study was supported by NIH grants R00 EB 009106, Discovery Foundation; Araclon Biotech; BioClinica, Inc.; to SS. The authors would like to thank Todd Peterson and Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.;
2084 submit your manuscript | www.dovepress.com Clinical Interventions in Aging 2017:12 Dovepress Dovepress Application of HFs in brain [18F]-florbetapir PET
Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly
RTL 0.583 0.584 0.576 0.57 0.578 0.562 0.537 and Company; EuroImmun; F Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE
LTL 0.602 0.546 0.601 0.618 0.636 0.601 0.502 Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Phar-
RPL 0.624 0.575 0.621 0.601 0.604 0.626 0.575 maceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics,
LPL 0.596 0.56 0.599 0.601 0.596 0.605 0.595 LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and RFL 0.589 0.504 0.591 0.583 0.587 0.591 0.517 Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites LFL 0.565 0.514 0.567 0.557 0.558 0.565 0.5 in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health (www.fnih. ADAS-Cog negative vs positive PCC 0.57 0.547 0.56 0.622 0.619 0.542 0.513 org). The grantee organization is the Northern California Institute for Research and Education, and the study is coor- RTL 0.555 0.589 0.557 0.665 0.657 0.568 0.624 dinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are LTL 0.548 0.591 0.556 0.667 0.645 0.571 0.57 disseminated by the Laboratory for Neuro Imaging at the University of Southern California.
RPL 0.558 0.629 0.559 0.576 0.575 0.56 0.522 Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI)
LPL 0.538 0.573 0.541 0.539 0.544 0.549 0.507 database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementa-
RFL 0.595 0.631 0.609 0.55 0.55 0.639 0.568 tion of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of
LFL 0.542 0.516 0.546 0.524 0.523 0.565 0.584 ADNI investigators can be found at: http://adni.loni.usc.edu/ wp-content/uploads/how_to_apply/ADNI_Acknowledge- ment_List.pdf. APOE carriers vs non-carriers PCC 0.569 0.547 0.578 0.622 0.621 0.594 0.608 Disclosure RTL 0.52 0.598 0.533 0.613 0.644 0.558 0.601 The authors report no conflicts of interest in this work.
LTL 0.536 0.613 0.546 0.579 0.619 0.565 0.562 References 1. Camus V, Payoux P, Barré L, et al. Using PET with 18F-AV-45 (florbeta- pir) to quantify brain amyloid load in a clinical environment. Eur J Nucl RPL 0.664 0.589 0.667 0.675 0.671 0.661 0.51 Med Mol Imaging. 2012;39(4):621–631. 2. Huang KL, Lin KJ, Hsiao T, et al. Regional amyloid deposition in amnestic mild cognitive impairment and Alzheimer’s disease evaluated LPL 0.621 0.55 0.621 0.69 0.686 0.624 0.53 by [18 F] AV-45 positron emission tomography in Chinese population. PLoS One. 2013;8(3):e58974. 3. Lopresti BJ, Klunk WE, Mathis CA, et al. Simplified quantification of RFL 0.61 0.595 0.612 0.647 0.639 0.614 0.548 Pittsburgh compound B amyloid imaging PET studies: a comparative analysis. J Nucl Med. 2005;46(12):1959–1972. 4. Ikonomovic MD, Klunk WE, Abrahamson EE, et al. Post-mortem LFL 0.626 0.603 0.623 0.615 0.603 0.616 0.564 correlates of in vivo PiB-PET amyloid imaging in a typical case of Alzheimer’s disease. Brain. 2008;131(6):1630–1645. 5. Landau SM, Fero A, Baker SL, et al. Measurement of longitudinal Normal vs MCI PCC 0.577 0.644 0.569 0.694 0.695 0.547 0.578 β-amyloid change with 18F-florbetapir PET and standardized uptake ADAS-Cog, Alzheimer’s Disease Assessment Scale-cognitive subscale; AUC, area under the curve; LFL, left frontal lobes; LPL, left parietal lobe; LTL, left temporal lobe; MCI, mild cognitive impairment; PCC, posterior value ratios. J Nucl Med. 2015;56(4):567–574.
AUC values derived from ROC curve analysis 6. Brendel M, Högenauer M, Delker A, et al; Alzheimer’s Disease
Blue highlight indicates AUCs above 0.6. Neuroimaging Initiative. Improved longitudinal [(18)F]-AV45 amyloid PET by white matter reference and VOI-based partial volume effect
Table 5 Feature Contrast Correlation Dissimilarity Energy Entropy Homogeneity S UV R Note: Abbreviations: cingulate cortex; RFL, right frontal lobes; ROC, receiver operating characteristic; RPL, parietal lobe; RTL, temporal SUVR, standardized uptake value ratio. correction. Neuroimage. 2015;108:450–459.
Clinical Interventions in Aging 2017:12 submit your manuscript | www.dovepress.com 2085 Dovepress Campbell et al Dovepress
7. Chen K, Roontiva A, Thiyyagura P, et al. Improved power for charac- 16. Wong DF, Rosenberg PB, Zhou Y, et al. In vivo imaging of amyloid terizing longitudinal amyloid-β PET changes and evaluating amyloid- deposition in Alzheimer disease using the radioligand 18F-AV-45 modifying treatments with a cerebral white matter reference region. (florbetapir [corrected] F 18).J Nucl Med. 2010;51(6):913–920. J Nucl Med. 2015;56(4):560–566. 17. Vanderbilt School of Engineering. Available from: http://masiweb. 8. Shokouhi S, Mckay JW, Baker SL, et al. Reference tissue normalization vuse.vanderbilt.edu/workshop2012/index.php/Main_Page. Accessed in longitudinal 18 F-florbetapir positron emission tomography of late June 6, 2017. mild cognitive impairment. Alzheimer’s Res Ther. 2016;8(1):2. 18. Open Access Series of Imaging Studies. Available from: http://www. 9. Matías-Guiu JA, Cabrera-Martín MN, Matías-Guiu J, et al. Amyloid oasis-brains.org/. Accessed June 6, 2017. PET imaging in multiple sclerosis: an 18 F-florbetaben study. BMC 19. Neuromorphometrics, Inc. Building a Model of the Living Human Neurol. 2015;15(1):243. Brain. Available from: http://neuromorphometrics.com/. Accessed 10. Englund E. Neuropathology of white matter changes in Alzheimer’s June 6, 2017. disease and vascular dementia. Dement Geriatr Cogn Disord. 1998; 20. Freeborough PA, Fox NC. MR image texture analysis applied to 9(Suppl 1):6–12. the diagnosis and tracking of Alzheimer’s disease. IEEE Trans Med 11. Head D, Buckner RL, Shimony JS, et al. Differential vulnerability of Imaging. 1998;17(3):475–478. anterior white matter in nondemented aging with minimal acceleration 21. Klyuzhin IS, Blinder S, Mabrouk R, Rahmim A, Sossi V. Investiga- in dementia of the Alzheimer type: evidence from diffusion tensor tion of texture quantification parameters for neurological PET image imaging. Cerebral Cortex. 2004;14(4):410–423. analysis. In: Nuclear Science Symposium and Medical Imaging Confer- 12. Charlton RA, Barrick TR, McIntyre DJ, et al. White matter damage on ence (NSS/MIC), 2015 IEEE; 2015:1–5. diffusion tensor imaging correlates with age-related cognitive decline. 22. Bruce CV, Clinton J, Gentleman SM, Roberts GW, Royston MC. Neurology. 2006;66(2):217–222. Quantifying the pattern of beta/A4 amyloid protein distribution in 13. Zhang Y, Schuff N, Du AT, et al. White matter damage in frontotem- Alzheimer’s disease by image analysis. Neuropathol Appl Neurobiol. poral dementia and Alzheimer’s disease measured by diffusion MRI. 1992;18(2):125–136. Brain. 2009;132(9):2579–2592. 23. Li Y, Tsui W, Rusinek H, et al. Cortical lamina binding of PET amyloid 14. Haralick RM, Shanmugam K. Textural features for image classification. and Tau Tracers in Alzheimer’s disease. J Nucl Med. 2015;56(2): IEEE Trans Syst Man Cybern. 1973;3(6):610–621. 270–273. 15. Rosen WG, Mohs RC, Davis KL. A new rating scale for Alzheimer’s disease. Am J Psychiatry. 1984;141(11):1356–1364.
Clinical Interventions in Aging Dovepress Publish your work in this journal Clinical Interventions in Aging is an international, peer-reviewed journal CAS, Scopus and the Elsevier Bibliographic databases. The manuscript focusing on evidence-based reports on the value or lack thereof of treatments management system is completely online and includes a very quick and fair intended to prevent or delay the onset of maladaptive correlates of aging peer-review system, which is all easy to use. Visit http://www.dovepress. in human beings. This journal is indexed on PubMed Central, MedLine, com/testimonials.php to read real quotes from published authors.
Submit your manuscript here: http://www.dovepress.com/clinical-interventions-in-aging-journal
2086 submit your manuscript | www.dovepress.com Clinical Interventions in Aging 2017:12 Dovepress